Title: Boosting CNN network performance for face recognition in an authentication system

Authors: Hamza Benyezza; Reda Kara; Mounir Bouhedda; Zine Eddine Safar Zitoun; Samia Rebouh

Addresses: Laboratory of Advanced Electronic Systems (LSEA), University of Medea, Nouveau Pôle Urbain, Medea, 26000, Algeria ' Laboratory of Advanced Electronic Systems (LSEA), University of Medea, Nouveau Pôle Urbain, Medea, 26000, Algeria ' Laboratory of Advanced Electronic Systems (LSEA), University of Medea, Nouveau Pôle Urbain, Medea, 26000, Algeria ' Laboratory of Advanced Electronic Systems (LSEA), University of Medea, Nouveau Pôle Urbain, Medea, 26000, Algeria ' Laboratory of Experimental Biology and Pharmacology (LBPE), University of Medea, Nouveau Pôle Urbain, Medea, 26000, Algeria

Abstract: Face recognition technology has made significant advancements through the utilisation of convolutional neural networks (CNN) in various applications. However, accurately identifying individuals from similar backgrounds remains a notable challenge due to inherent similarities in facial features among individuals with shared genetic ancestry or cultural heritage. This paper addresses the limitations of traditional CNN in accurately identifying individuals from the same origins and presents an approach to enhance the performance of CNN networks and improve the reliability of face recognition in authentication systems. The proposed approach incorporates an advanced face detection and identification algorithm based on the visual geometry group face (VGG-Face) CNN descriptor model, along with the cosine distance algorithm. Promising results were obtained through a prototype implementation on a Raspberry Pi 4. Comparative evaluations against alternative face recognition strategies showcased exceptional performance, achieving an accuracy rate of 96.33% for positive pairs and 95.38% for negative pairs at an optimal threshold of 20.

Keywords: smart authentication system; face detection and identification; VGG-face CNN descriptor; Internet of Things; cosine distance algorithm.

DOI: 10.1504/IJDATS.2024.140648

International Journal of Data Analysis Techniques and Strategies, 2024 Vol.16 No.3, pp.282 - 310

Received: 06 Jun 2023
Accepted: 02 Mar 2024
Published online: 29 Aug 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article